Evaluating The Embedding Space of Foundation Models for Dermatological Images to Guide Backbone Selection for A Fine-Tuning Pipeline

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Thi Anh Thu Pham, Khoi Nguyen Gia, Tran Hieu, Tran Anh Duc, Thi Thanh Tu Bui
Format: Recurso digital
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901268794966016
author Thi Anh Thu Pham
Khoi Nguyen Gia
Tran Hieu
Tran Anh Duc
Thi Thanh Tu Bui
author_facet Thi Anh Thu Pham
Khoi Nguyen Gia
Tran Hieu
Tran Anh Duc
Thi Thanh Tu Bui
contents In recent years, computer vision models have advanced rapidly and have been increasingly applied in dermatology. However, in practice, many skin diseases present with very similar visual manifestations, making image representation in vector form not always accurately reflect the true nature of the disease. Therefore, this study focuses on comparing two approaches: general-purpose models designed for multiple tasks and models specifically trained for medical and dermatological domains.Experiments were conducted on 1,300 dermatological images covering 13 different infectious diseases, with a balanced number of images across disease categories. The models were evaluated based on their ability to retrieve images with similar characteristics, while also considering computational cost.The results show that in terms of pure representation quality, an ultra-large-scale general-purpose model achieved the highest performance. However, dermatology-specific models provided an optimal balance between performance and computational cost for the infectious disease group, producing stable and clinically appropriate image representations to serve as a foundation for subsequent clinical research.
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19388449
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Evaluating The Embedding Space of Foundation Models for Dermatological Images to Guide Backbone Selection for A Fine-Tuning Pipeline
Thi Anh Thu Pham
Khoi Nguyen Gia
Tran Hieu
Tran Anh Duc
Thi Thanh Tu Bui
Dermatology Embedding; Embedding Evaluation; Foundation Models; Backbone Selection; Skin Retrieval
In recent years, computer vision models have advanced rapidly and have been increasingly applied in dermatology. However, in practice, many skin diseases present with very similar visual manifestations, making image representation in vector form not always accurately reflect the true nature of the disease. Therefore, this study focuses on comparing two approaches: general-purpose models designed for multiple tasks and models specifically trained for medical and dermatological domains.Experiments were conducted on 1,300 dermatological images covering 13 different infectious diseases, with a balanced number of images across disease categories. The models were evaluated based on their ability to retrieve images with similar characteristics, while also considering computational cost.The results show that in terms of pure representation quality, an ultra-large-scale general-purpose model achieved the highest performance. However, dermatology-specific models provided an optimal balance between performance and computational cost for the infectious disease group, producing stable and clinically appropriate image representations to serve as a foundation for subsequent clinical research.
title Evaluating The Embedding Space of Foundation Models for Dermatological Images to Guide Backbone Selection for A Fine-Tuning Pipeline
topic Dermatology Embedding; Embedding Evaluation; Foundation Models; Backbone Selection; Skin Retrieval
url https://doi.org/10.5281/zenodo.19388449